A 2024 arXiv study (Yang) manually built the directed following network among 65 active US news accounts on Twitter/X, colored the nodes with six-level AllSides media-bias labels, and read Gephi community statistics (modularity 0.16, three ideologically aligned communities, clustering coefficient 0.3) as evidence of an echo chamber among news outlets. The study's own limitations section states the unresolved question: whether this outlet-level echo chamber is an artifact of a narrow measurement design — outlet-only nodes, a single tie type (following), unweighted edges, and third-party bias labels validated only once — and how outlet-level clustering relates to what audiences are actually exposed to. The open problem is to measure the US news ecosystem on Twitter/X with audience-inclusive node sets, multiple interaction modalities, edge-weighting schemes, and validated bias labels, and to determine whether the echo-chamber conclusion and its magnitude survive, and how outlet-level segregation translates into audience-level exposure. Prior literature answers the components in general form and in other contexts (retweet-vs-mention segregation, follow-vs-retweet homophily, individual-level exposure studies, LLM-assisted bias labeling), and the field-wide debate over echo-chamber existence and prevalence remains method-divided; no study has performed the integrated robustness-and-exposure analysis for this ecosystem.

Background

An 'echo chamber' is a socially homogeneous information environment in which people encounter predominantly like-minded news and opinion, reinforcing prior beliefs. Whether social media platforms create echo chambers — with what prevalence and magnitude — is one of the most contested questions in political communication, with implications for democratic debate and platform regulation. The US news ecosystem on Twitter/X is a central battleground for this debate because its interaction structure is observable: which accounts follow, retweet, like, or mention which others can be recorded as a network and analyzed with community-detection tools.

Network-analytic studies represent actors as nodes and one tie type as edges, then measure homophily (the tendency of similar nodes to connect) and modularity (the strength of a partition of the network into communities, computed for example by the Gephi software; higher modularity means denser within-community than between-community ties). A recurring finding is that conclusions depend on these design choices. Conover et al. (2011) showed that the political retweet network on Twitter is sharply segregated by ideology while the mention network is not, and Colleoni et al. (2014) found that political homophily differs between the follow and retweet networks — the chosen tie type can flip the qualitative conclusion. Audience-level studies paint an equally split picture. Bakshy, Messing and Adamic (2015) measured individual exposure to cross-cutting news on Facebook and found individual choice matters more than the ranking algorithm. Eady et al. (2019) linked a representative survey to respondents' Twitter follows and found substantial overlap in the ideological distributions seen by left- and right-leaning users, though with asymmetries. Wojcieszak et al. (2022) found that most Twitter users follow no political elites at all, while those who do favor ideologically congenial accounts roughly nine-to-one. Gentzkow and Shapiro (2011) found ideological segregation of news consumption modest overall, offline and online. At the user level, Cinelli et al. (2021) documented homophilic clustering across platforms, while Terren and Borge-Bravo's (2021) systematic review of 55 studies concluded that digital-trace studies tend to find echo chambers and self-report studies tend not to — the method is itself part of the controversy. The measurement frontier keeps moving: Pratelli, Saracco and Petrocchi (2024) proposed an entropy-based echo-chamber detector and found that only about 0.35% of users in an Italian Twitter debate occupy echo chambers, and Di Martino et al. (2025) reframed fragmentation at the level of whole platforms ('echo platforms').

Within this landscape, Wen Yang's 2024 preprint took a deliberately small, outlet-level slice of the question. Using bias labels from AllSides — a service that rates hundreds of US outlets on a six-level scale from 'left' through 'center' to 'right', plus a distinct 'allsides' category for outlets that deliberately publish multiple perspectives — the preprint manually extracted the directed following relation among 65 active official news accounts on Twitter/X (activity threshold: combined follower-plus-following count above 5, chosen because free API access no longer permitted automated crawling) and analyzed the resulting binary adjacency matrix. The preprint describes its data only in prose (a bias-label file and a 65x65 following matrix) and does not publish them. Gephi statistics (average degree 10.2, density 0.16, clustering coefficient 0.3, modularity 0.16 with three communities matching the left/center/right labels) and network visualizations were read as evidence of an echo chamber at the level of news outlets. The bias labels were validated only by a single correlation (0.454) with a text-modeling slant measure in the tradition of Gentzkow and Shapiro (2010), and the author acknowledged possible label bias for right-leaning accounts. The preprint's own limitations section states what remains unresolved: the node set excludes audiences; only the following relation was measured (posts, reposts, likes, comments were not); the edges are unweighted even though a weighting scheme that scores within-category follows lower and dissimilar-category follows higher could change the network statistics and the echo-chamber conclusion; bias labels await automated validation and refinement; and the manual collection does not scale to larger samples.

Problem Statement

Taking Yang (2024) as the baseline — the directed, unweighted following network among 65 active US news accounts on Twitter/X, nodes labeled with six-level AllSides bias ratings, in which modularity-based community detection yields three ideologically aligned communities interpreted as an echo chamber — determine whether that echo-chamber conclusion is an artifact of the measurement design, and establish how outlet-level clustering relates to audience-level exposure. Any enriched measurement must be compared against a network built under the source's design whose baseline network statistics are comparable to the source's reported values, so that any change in structure can be attributed to the enriched measurement rather than to a different starting network. Concretely, using the source's design as the reference point: (1) extend the node set beyond outlets to include audience/user nodes and their exposure, so that outlet-outlet segregation can be compared with what users actually follow and see; (2) incorporate interaction modalities beyond the following relation (posts, reposts/retweets, likes, comments, mentions) and test whether community structure and clustering statistics change across tie types; (3) test sensitivity of the detected structure to edge-weighting schemes, including schemes that assign lower weight to within-category followings and higher weight to dissimilar-category followings, as the source proposes; (4) validate or refine the bias labels with automated content analysis (for example natural language processing (NLP) or large-language-model (LLM) methods), quantifying label error and its downstream effect on the detected communities; and (5) scale the data collection beyond the manually extracted 65 nodes, including larger outlet samples. Taken together, these design-dimension checks bear on one question: whether, and with what magnitude, echo chambers in the US news ecosystem on Twitter/X persist under enriched measurement, and how outlet-level clustering translates into audience exposure. The audience-inclusive exposure mapping (1), the tie-type comparison (2), the edge-weighting sensitivity (3), and the bias-label refinement together with its effect on the detected communities (4) each address a distinct way the echo-chamber conclusion could fail, and outlet-level clustering must be read together with audience exposure; scaling the collection beyond the manually extracted 65 nodes (5) is a further extension of the same design.

References

  1. Wen Yang (2024). Are There Echo Chambers in the US News Ecosystem? Evidence From Twitter/X. arXiv:2404.15631. https://arxiv.org/abs/2404.15631
  2. Michael D. Conover, Jacob Ratkiewicz, Matthew Francisco, Bruno Goncalves, Filippo Menczer, Alessandro Flammini (2011). Political Polarization on Twitter. Proceedings of the International AAAI Conference on Web and Social Media 5(1). DOI: 10.1609/icwsm.v5i1.14126. https://doi.org/10.1609/icwsm.v5i1.14126
  3. Elanor Colleoni, Alessandro Rozza, Adam Arvidsson (2014). Echo Chamber or Public Sphere? Predicting Political Orientation and Measuring Political Homophily in Twitter Using Big Data. Journal of Communication 64(2):317-332. DOI: 10.1111/jcom.12084. https://doi.org/10.1111/jcom.12084
  4. Matteo Cinelli, Gianmarco De Francisci Morales, Alessandro Galeazzi, Walter Quattrociocchi, Michele Starnini (2021). The echo chamber effect on social media. Proceedings of the National Academy of Sciences 118(9):e2023301118. DOI: 10.1073/pnas.2023301118. https://doi.org/10.1073/pnas.2023301118
  5. Ludovic Terren, Rosa Borge-Bravo (2021). Echo Chambers on Social Media: A Systematic Review of the Literature. Review of Communication Research 9:99-118. DOI: 10.12840/ISSN.2255-4165.028. https://doi.org/10.12840/ISSN.2255-4165.028
  6. Eytan Bakshy, Solomon Messing, Lada A. Adamic (2015). Exposure to ideologically diverse news and opinion on Facebook. Science 348(6239):1130-1132. DOI: 10.1126/science.aaa1160. https://doi.org/10.1126/science.aaa1160
  7. Gregory Eady, Jonathan Nagler, Andy Guess, Jan Zilinsky, Joshua A. Tucker (2019). How Many People Live in Political Bubbles on Social Media? Evidence From Linked Survey and Twitter Data. SAGE Open 9(1). DOI: 10.1177/2158244019832705. https://doi.org/10.1177/2158244019832705
  8. Magdalena Wojcieszak, Andreu Casas, Xudong Yu, Jonathan Nagler, Joshua A. Tucker (2022). Most users do not follow political elites on Twitter; those who do show overwhelming preferences for ideological congruity. Science Advances 8(39):eabn9418. DOI: 10.1126/sciadv.abn9418. https://doi.org/10.1126/sciadv.abn9418
  9. Matthew Gentzkow, Jesse M. Shapiro (2011). Ideological Segregation Online and Offline. The Quarterly Journal of Economics 126(4):1799-1839. DOI: 10.1093/qje/qjr044. https://doi.org/10.1093/qje/qjr044
  10. Jesse Shore, Jiye Baek, Chrysanthos Dellarocas (2018). Network Structure and Patterns of Information Diversity on Twitter. MIS Quarterly 42(3):849-872. DOI: 10.25300/misq/2018/14558. https://doi.org/10.25300/misq/2018/14558
  11. Matthew Gentzkow, Jesse M. Shapiro (2010). What Drives Media Slant? Evidence From U.S. Daily Newspapers. Econometrica 78(1):35-71. DOI: 10.3982/ECTA7195. https://doi.org/10.3982/ECTA7195
  12. Raphael Hernandes, Giulio Corsi (2024). LLMs left, right, and center: Assessing GPT's capabilities to label political bias from web domains. arXiv:2407.14344. https://arxiv.org/abs/2407.14344
  13. Manuel Pratelli, Fabio Saracco, Marinella Petrocchi (2024). Entropy-based detection of Twitter echo chambers. PNAS Nexus 3(5):pgae177. DOI: 10.1093/pnasnexus/pgae177. https://doi.org/10.1093/pnasnexus/pgae177
  14. Edoardo Di Martino, Alessandro Galeazzi, Michele Starnini, Walter Quattrociocchi, Matteo Cinelli (2025). Ideological fragmentation of the social media ecosystem: From echo chambers to echo platforms. PNAS Nexus. DOI: 10.1093/pnasnexus/pgaf262. https://doi.org/10.1093/pnasnexus/pgaf262
  15. Pawel Matuszewski, Gabriella Szabo (2019). Are Echo Chambers Based on Partisanship? Twitter and Political Polarity in Poland and Hungary. Social Media + Society 5(2). DOI: 10.1177/2056305119837671. https://doi.org/10.1177/2056305119837671
  16. Ardian Maulana, Hokky Situngkir (2020). Media Polarization on Twitter During 2019 Indonesian Election. In: Complex Networks & Their Applications IX, Studies in Computational Intelligence, 660-670. DOI: 10.1007/978-3-030-65347-7_55. https://doi.org/10.1007/978-3-030-65347-7_55
  17. Kyle Rose, Deana A. Rohlinger (2024). Political Influencers and Their Social Media Audiences during the 2021 Arizona Audit. Socius 10. DOI: 10.1177/23780231241259680. https://doi.org/10.1177/23780231241259680
  18. Gregory Eady, Richard Bonneau, Joshua A. Tucker, Jonathan Nagler (2024). News Sharing on Social Media: Mapping the Ideology of News Media, Politicians, and the Mass Public. Political Analysis 33:73-90. DOI: 10.1017/pan.2024.19. https://doi.org/10.1017/pan.2024.19